Engineering logbook

8 min· sqlite· browser· performance

Why I put SQLite in the browser: IndexedDB can't filter on one field and sort by another

The one query that made me stop working around IndexedDB and start replacing it. What a cursor can and cannot do, why fetch-everything-and-sort-in-JS stops working at a few thousand rows, and what a WASM SQLite build actually costs you.

8 min· browser· architecture· sqlite

When a browser app needs a real database — and when it absolutely doesn't

Four symptoms that mean your client-side storage has outgrown localStorage, four situations where adding a database makes things worse, and the measurement that decides it. Written by someone who builds one of these, trying hard to talk you out of it.

15 min· postmortem· debugging· sqlite

The bug that stored your file as {}: typeof x === 'object' is not a plain object

I wrote a codec specifically to stop JSON silently mangling values, and it destroyed every type I forgot to list. Blob, ArrayBuffer, Map, Set and RegExp all stored as an empty object, with nothing thrown. Why serialisers fail in the direction of silence, and the three-part fix where missing one part leaves it broken.

9 min· sqlite· browser· concurrency

Two tabs, one SQLite file: the hard part of a browser database

OPFS sync access handles are exclusive, so two tabs writing one database file is corruption, not a race you retry. How single-writer election over Web Locks works, why the naive VFS needs COOP/COEP you cannot ship, and the two failure modes that look identical and must be handled differently.

measured · aug 08

macOS said 150 GB of Documents. The folder was 6.4 MB.

Mac storage settings blamed 'Documents' for over 150 GB. The real consumers were 43 GB of node_modules and .next across 77 git worktrees and 31 GB of Docker cache. The commands that found it, and the three fixes that keep it from coming back.

7 min· macos· disk space· node_modules

shipped · aug 08

Deleting git worktrees safely: my first two merge checks were wrong

Building a CLI that deletes git worktrees means answering 'is this branch merged?' when your platform squash-merges everything. git branch --merged can't. git cherry only half can. Two real defects, and the tests that passed while protecting nothing.

11 min· git· cli· testing

field_note · aug 08

I audited a 185-table Postgres database: ACID, normalization and deadlocks

22 findings from auditing a live Postgres application: a transaction that commits on the way out, a rescue that catches nothing once the transaction is aborted, derived values with no constraint behind them, a currency filter that zeroes a rollup, and a row-level deadlock that took four attempts to fix. Every claim has the reproduction that proved it.

28 min· postgres· acid· database design

field_note · aug 05

Testing webhooks locally, in both directions

A webhook integration has two directions and each fails in its own way. How to drive both from your laptop — signing requests the way the receiver verifies them, proving every event type actually fires, and asserting the async tail — plus the guard that makes your own machine unreachable by design.

13 min· webhooks· testing· integration

field_note · aug 05

What I found auditing 79 background jobs

A Postgres-backed job queue that looked fine had a concurrency guard that permitted the thing it named, transactions held open across SMTP calls, and sweeps that lost a night's work to one bad row. Seven lessons, with the code, and the verification discipline that caught my own wrong answers.

17 min· rails· postgres· background jobs

field_note · aug 01

Your per-user rate limit is probably counting IPs

Rate limiting runs in middleware. Authentication usually runs in the controller. If your limiter asks who the user is, it asks before anyone has answered — and every per-user bucket silently collapses into one bucket per IP.

7 min· rate limiting· rails· middleware

field_note · aug 01

Migrating a live server across CPU architectures: the seven things that weren't in my plan

Moving a single-host production service from x86 to ARM, and rewriting its storage layer the same night. Why you can't resize across architectures, why grep can't find every reference to a host, the two-pass rsync that shrank the outage to five seconds, and the checksum that proved nothing was lost.

19 min· infrastructure· terraform· aws

field_note · jul 31

Your invite endpoint is a mail relay

Three throwaway signups sent 127,158 phishing emails through a B2B SaaS team-invite feature, with the scam text carried in the workspace name field. Every control that should have stopped it existed, reported healthy, and could not fire.

10 min· security· postmortem· rate limiting

10 min· observability· sre· rails

Your health check is lying to you: liveness for workers, proxies and the edge

A 200 from /up proves one web process answered one request. It says nothing about background workers, and nothing at all about whether users can reach you. Four signals that close the gap, with the Rails and Python code.

5 min· terraform· aws· iam

Two Terraform configs, one live resource: how a tag became an access-control bug

A pre-refactor Terraform directory nobody ran still described the production database — with different tags. Applying it would have stripped the tag that gates SSO access and groups the cost report. Why tag drift is a security finding, not a tidiness one.

8 min· observability· cloudwatch· sre

The alarm that pages on its first day — and the fix that broke reconciliation

A brand-new CloudWatch alarm fired critical while the endpoint it watched was serving 200s in 100ms. The cause is a documented CloudWatch behaviour almost nobody reads, and the guard I added for it quietly broke the drift detection two days later.

11 min· observability· aws· cloudwatch

AWS alerts into a chat channel, without a monitoring vendor

A production account with zero CloudWatch alarms, wired end to end in a weekend: CloudWatch and RDS events into SNS, one stdlib Lambda into Discord, and a committed JSON file as the source of truth. Full code, and the four design decisions that matter.

8 min· observability· sre· cloudwatch

Stop guessing alarm thresholds: measure them, then replay an outage

Round numbers make alarms that flap or never fire. A 120-line baseline script, two thresholds corrected before rollout, and the validation step almost nobody does — replaying a real incident window to prove the alarm would have fired in time.

postmortem · jul 29

I profiled the wrong database for an hour

A login timing out at 166 seconds on an idle box. The cause was a burstable RDS instance out of CPU credits — but a stale comment sent me profiling a database the app doesn't use, and log retention had already deleted the evidence. The trace, the two wrong turns, and what I'd instrument first next time.

8 min· debugging· postgresql· aws

field_note · jul 29

The complete ML pipeline: from raw data to production, with the calculations

Every stage of shipping a machine learning model — data cleaning, splits, tokenisers, metrics, training maths, evaluation design, quantisation, serving capacity and deployment — with the formulas worked through and the numbers that decide each choice.

21 min· machine learning· mlops· ml engineering

field_note · jul 29

The ASR heard English perfectly and wrote it in the wrong alphabet

A fine-tuned speech model transcribed accented English into Devanagari with 0.000 script accuracy. Fixing it meant training on eighteen accents that weren't the target one — plus a tokenizer bug that cost a nine-hour run.

12 min· machine learning· asr· fine-tuning

shipped · jul 28

Nepali speech-to-text: why AI ignores your voice notes

The best free model wasn't mishearing Nepali — it was writing it in the wrong alphabet. The diagnosis, the fine-tune that fixed it, and what it cost.

7 min· speech recognition· ai· self-hosted

field_note · jul 28

Your rate limiter's fixed window is a 2× burst

A fixed-window counter lets a client spend two full budgets inside two minutes, and the obvious Postgres implementation of a token bucket silently loses writes under concurrency. Both fixes, with the SQL.

7 min· postgres· rate limiting· concurrency

shipped · jul 28

Self-hosted speech-to-text API: the whole pipeline

An OpenAI-compatible endpoint, an int8 ONNX model, a container that won't ship the wrong weights, and the queue isolation that keeps long clips honest.

9 min· self-hosted· api design· onnx

field_note · jul 28

Postmortem: 487,000 webhooks in eleven hours

A partner platform's replay loop delivered half a million events into a CRM's webhook receiver in one working day and pinned the box at 99% CPU. The investigation, the root cause, and why every existing guard was the wrong shape.

9 min· postmortem· webhooks· rate limiting

measured · jul 28

MLOps for fine-tuning a speech model on a free GPU

Eighteen failed attempts, then one clean run. What separated them: reproducible splits, an eval that measures the shipping code, and a free tier as spec.

7 min· mlops· fine-tuning· reproducibility

field_note · jul 28

Debugging a box you can only reach through SSM

No SSH, no bastion, no open ports, and an IAM role that can't even SendCommand. Investigating a webhook flood through Session Manager and docker exec — with every command, and the four ways it silently lies to you.

8 min· aws· docker· ssm

postmortem · jul 28

Catastrophic forgetting: what fine-tuning really costs

A fine-tune halved error on one language and broke two others. Then the textbook fix measured worse. Three failures that all looked green until they didn't.

8 min· postmortem· fine-tuning· machine learning

field_note · jul 25

Your AI coding agent doesn't know your laptop is on fire

AI coding agents fire off make -j, parallel test suites, and cargo build with no idea your machine is already pinned. Here's heatsink — a tiny open-source guard that right-sizes runaway parallel commands before they cook your laptop.

7 min· ai· developer-tools· open-source

WebMCP: letting an in-browser AI agent call your tools instead of clicking them

Browser agents drive your site by screenshotting and guessing where to click. WebMCP (navigator.modelContext) lets the page hand the agent typed, callable tools instead. Why it's nearly free if your tools are already client-side, and where it still falls short.

4 min· ai· mcp· agents

4 min· ocr· browser extension· productivity

The browser extension that replaces a dozen clunky OCR tools: inside Textquill

Most people have a graveyard of half-used OCR apps, screenshot uploaders, and paste-into-a-website tools. A single on-device browser extension collapses the whole pile. Here's the case for consolidating on Textquill.

4 min· ocr· privacy· browser extension

From screenshots to editable text in seconds: how Textquill keeps your data truly private

The gap between 'we don't store your images' and 'your images never left your device' is the whole ballgame. Here's how Textquill turns screenshots into editable text without a server ever seeing them.

5 min· cloudflare· deployment· operations

One box, zero open ports, no Elastic IP: deploying a production service behind a Cloudflare Tunnel

A real single-box deployment where the firewall blocks all inbound, there's no public IP, and no Elastic IP on the bill. How Cloudflare Tunnel becomes the only ingress, how the containers are deployed and networked, and what it saves in operations and cost.

4 min· ocr· privacy· browser extension

Why on-device OCR is the smarter way to extract text: a look at Textquill

Cloud OCR sends every screenshot to someone else's server. On-device OCR runs the model where the pixels already are. Here's why that trade keeps winning, using Textquill as the example.

4 min· devops· kamal· terraform

Shipping to one box, safely: Kamal + Terraform + GitHub OIDC with no long-lived AWS keys

The full deploy pipeline for a single-box production service — Terraform provisions the host, network, IAM and edge; GitHub Actions federates into AWS via OIDC (no stored keys); CI gates on security scans; Kamal ships containers over SSH-through-SSM with zero open ports.

postmortem · jul 19

"NetworkError when attempting to fetch": debugging a Cloudflare 1033 tunnel flap

A login page throwing NetworkError led through Cloudflare error 1033, a cloudflared QUIC transport that silently stalled, and a Kamal gotcha that made the first fix look deployed when it wasn't. The trace, the permanent fix (http2), and an auto-heal watchdog.

5 min· debugging· cloudflare· operations

4 min· terraform· cloudflare· infrastructure

cloudflared tunnel token in Terraform v5: the tunnel_secret example that actually works

The cloudflare_zero_trust_tunnel_cloudflared_token data source returns a token, not tunnel_secret. Here's the v5 schema, where tunnel_secret really lives, and a full working example.

field_note · jul 18

Screenshot to text: turn any screen region into editable text

How to go from a screenshot, or any region of your screen, to editable text in one action, without saving a file or uploading anything. A practical guide for developers, students, and researchers.

2 min· ocr· screenshot· chrome-extension

field_note · jul 18

Private OCR: extract text from images without uploading them

Most OCR tools send your image to a server. Here's why that's a problem for anything sensitive, and how on-device OCR in the browser reads text without the image ever leaving your machine.

2 min· ocr· privacy· chrome-extension

field_note · jul 18

How to extract text from an image in Chrome (free, no upload)

Four ways to pull editable text out of an image in Chrome, from copy-typing to cloud OCR sites to a free on-device extension, and an honest look at which one you'll actually keep using.

3 min· ocr· chrome-extension· productivity

shipped · jul 12

The AI prompt manager I actually use every day

Why I stopped keeping prompts in Notes and built Prompt Vault — a free Chrome extension that saves, organizes, and one-click inserts prompts into ChatGPT, Claude, and Gemini.

2 min· ai· prompts· productivity

field_note · jul 11

How to save your ChatGPT prompts (and actually find them again)

Three ways to save your ChatGPT prompts — notes, spreadsheets, and a dedicated prompt manager — and an honest look at which one you'll still be using in a month.

2 min· ai· prompts· chatgpt

field_note · jul 10

Reusable prompt templates: stop rewriting the same prompt

Turn your best prompts into templates with variables so one prompt covers a hundred situations — with a simple pattern and a free tool to run it in ChatGPT, Claude, and Gemini.

2 min· ai· prompts· templates

field_note · jul 09

The best AI prompt manager in 2026: what to actually look for

Most 'best AI prompt manager' lists rank features nobody uses. Here are the five that decide whether you'll still be using the tool in a month — and the one I built around them.

2 min· ai· prompts· tools

field_note · jul 08

How to organize a prompt library that doesn't become a junk drawer

A simple folder, tag, and naming system for organizing your AI prompts so you can find any of them in a keystroke — even at two hundred prompts.

2 min· ai· prompts· organization

shipped · jun 26

Drydock: from one coding agent to a 19-agent product team

AI writes the code — the easy 10%. Drydock is an open-source agent team for Claude Code that runs the rest, idea to launch: architecture, tests, security, CI/CD, compliance, and go-to-market.

4 min· ai· agents· claude

field_note · jun 23

Cloudflare Tunnel Docker Compose: Zero-Port Stack Setup

Set up Cloudflare Tunnel Docker Compose with zero inbound ports. One container routes multiple services via Docker's internal DNS. Full setup guide.

10 min

field_note · jun 23

Cloudflare Tunnel No Open Ports Setup: Verify Zero Exposure

Eliminate inbound firewall rules by configuring Cloudflare Tunnel with zero open ports. Learn the full setup with ss and external nmap proof steps.

11 min

field_note · jun 23

Reduce Claude AI Hallucinations in Agentic Workflows

Reduce Claude AI hallucinations in production pipelines with RAG, temperature tuning, forced tool use, and validation. Five concrete API-level techniques.

14 min

Claude Code Agentic AI: Deploy an Agent in Minutes

Learn to build and deploy a Claude Code agent in under 15 minutes. Real wiring, tool bindings, costs, and two gotchas that kill most first runs. Start now.

11 min

Master All 6 Claude Code Workflows: Decision Guide

Master all 6 Claude Code workflows: decision matrix by scope, reversibility, and risk. Compare token costs, latency, and get ready-to-use CLAUDE.md templates.

16 min

Loop Engineering: Designing Systems That Prompt AI Agents

Loop engineering: master state transitions, feedback injection, termination contracts, and failure recovery. The system design discipline that makes AI agents

15 min